just use openrouter or google ai playground for the first week till bugs are ironed out. You still learn the nuances of the model and then yuu can switch to local. In addition you might pickup enough nuance to see if quantization is having any effect

Yes, exactly. I like this analogy. I am surprised the level of pearl clutching in these discussions on Hacker News. Everybody wants to be an attention sharecropper, lol.

unconstrained AI agents are what makes it so useful though. I have been using claude for almost a year now and the biggest unlock was to stop being a worrywart early on and just literally giving it ssh keys and telling it to fix something. ofc I have backups and do run it in VM but in that VM it helps me manage by infra and i have a decent size homelab that would be no fun but a chore without this assistant.

I run my AI agent unconstrained in a VM without access to my local network so it can futz with the system however it wants (so far, I've had to rebuild the VM twice from Claude borking it). That works great for software development.

For devops work, etc (like your use case), I much prefer talking to it and letting it guide me into fixing the issue. Mostly because after that I really understand what the issue was and can fix it myself in the future.

Letting an agent loose with SSH keys is fine when the blast radius is one disposable VM, but scale that habit to prod or the wrong subnet and you get a fast refresher on why RBAC exists, why scoped creds exist, and why people who clean up after outages get very annoyed by this whole genre of demo. Feels great, until it doesn't.

jai is doing the right thing for its threat model. The credential layer is a different surface though ... an agent with a broad API token can call initiate_payment or update_vendor_bank on a remote production system and the filesystem sandbox can't help.

Applying the same principle as jai for remote boundaries, we can scope API authority to the task

Agree, but SSH agents like 1Passwords are nice for that.

You simply tell it to install that Docker image on your NAS like normal, but when it needs to login to SSH it prompts for fingerprint. The agent never gets access to your SSH key.

> unconstrained AI agents are what makes it so useful though

Not remotely worth it.

It's not really malware, but it's a mess. It installed so much shit and it interfered with your git hooks and stuff. It was kind of messy. I kind of gave up on it. I just went back to using built-in claude code todowrite tasks.

It managed to throw itself into a global file for me that Claude used which caused beads to appear in random projects on my machine. Because of how it was there the agent attempted to re-install beads after I already removed it because the guy hook errored.

Haven't tried it, but this rewrite might be better?

https://github.com/Dicklesworthstone/beads_rust

Try https://github.com/hmans/beans - I find it a refreshingly pragmatic take that works great with my agents use.

I have run technitium for 4 or so years now, in a recursive mode, handles all my homelab needs and it is faster as well. Now that it has clustering support I have three instances in my proxmox cluster.

Do you need the simcard for WhatsApp?

I believe you only need a unique phone number to create the account, then you can use WhatsApp Web as client. Be very careful with alternative clients, as I've had an account banned in the past for this (and therefore a phone number blacklisted), even without messaging anybody. I think that clients that run WhatsApp Web in a web view (like https://github.com/rafatosta/zapzap) are safe.

I think they started banning unauthorized API users around the time that "WhatsApp For Business" was introduced, because it was competing with that product. Unfortunately WhatsApp For Business is geared toward physical products and services with registered companies, so home automation and agents are left with no options.

I believe you can use a virtual number/VOIP (like Twilio or Google Voice), but I want to be able to eventually use SMS where WhatsApp can't be used, so I do know some services identify "non residential" SMS phone numbers (for example I've seen Google Voice numbers blocked) so I wanted to prevent that from happen. Again, key thing here for me is that my assistant appears to be a human.

Great job. I had to disable that logi sw software on my Macbook because it was taking up a lot of cycles.

Exactly. Look at just the most recent conflict in Middle East. You think they would have freaking gamed out potential scenarios using AI or whatnot? Looks like nobody gamed out anything. It's all just seat of the pants.

All the people who had the job title of "War game Iran scenarios" were actually fired a month or so ago.

Competence in a senior position is a threat to an incompetocracy. It's more important you be stupid and loyal than be good.

The military has performed countless simulations and “what-if” exercises and thoroughly documented each one. They knew a war with Iran without boots on the ground doesn’t end with a decisive victory. Trump chose to ignore them and press ahead anyway.

You can’t really understand Trump’s decisions unless you understand that despite all evidence to the contrary, Trump himself truly believes he is the smartest person in the room, regardless of who else is in it; and he will not suffer anyone who dares to contradict him.

>Trump himself truly believes he is the smartest person in the room, regardless of who else is in it; and he will not suffer anyone who dares to contradict him.

I actually believe he has a crippling inferiority complex, which is why he leans so hard into bluster and bravado, why he surrounds himself with incompetent sycophants, and also why he's so vicious at even a hint of being slighted.

I think he probably knows, deep down, that he's mid at best and his most deep-seated fear is being perceived as insufficiently masculine, intelligent, powerful, wealthy, etc.

The fact that they did is likely why Trump fired one of his generals.

Ive worked in organizations like that where EVERYBODY knew something was a bad idea but upper management wanted to do it anyway. At some point you get frozen out if you dissent and nobody gives two halfs of a fuck about when it turns out you were right. Conformity is all that matters.

Even so a few people do publicly dissent.

I was building awesome things with Access 20 years ago. I loved that thing. I wasn't even a software engineer. I was in the EE, but I needed a way to track process and it definitely outperformed. And the best thing, it didn't cost us anything. Everybody already had access, lol. I had 40 people use it in production, manufacturing cutting edge stuff. Definitely beat spreadsheets because Access gave you gui for operators.

Q4 quants on 32G VRAM gives you 131K context for 35BA3B and 27B models who are pretty capable. On 5090 one gets 175 tg and ~7K pp with 35BA3B, 27B isaround 90 tg. So speed is awesome. Even Strix 395 gives 40 tk/s and 256K context. Pretty amazing, there is a reason people are excited about qwen 3.5